Supervized mixture of PLDA models for cross-channel speaker verification

Konstantin Simonchik, Timur Pekhovsky, Andrey Shulipa, Anton AfanasyevInterspeech 2012

TL;DR

This paper tackles the challenge of speaker verification across different audio channels (like phone calls vs. microphone recordings) by proposing a novel approach: a supervised mixture of Probabilistic Linear Discriminant Analysis (PLDA) models. This method significantly improves robustness and accuracy compared to individual channel-specific models, offering a more stable and reliable solution for real-world applications.

What Problem Does This Paper Address?

Speaker verification systems often struggle when the enrollment audio (what the system learns about a speaker) and the test audio (what the system verifies against) come from different sources. This "cross-channel" problem is particularly pronounced with telephone speech versus microphone recordings due to variations in microphones, transmission lines, and acoustic environments. Traditional methods, while improving, still face challenges in handling these significant channel differences. The paper builds upon previous work that used PLDA, a powerful technique for modeling speaker variability, and specifically addresses the need for robust cross-channel verification by exploring a mixture of models trained on different channel types.

Methods

The core of this paper is the development and application of a supervised mixture of PLDA models. Here's a breakdown:

  • PLDA Basics: PLDA models are used to represent speaker identity and channel effects in a low-dimensional "i-vector" space. The paper uses a standard PLDA formulation with Gaussian priors for speaker and channel factors.
  • Mixture of Models: Instead of a single PLDA model, the authors propose combining multiple PLDA models. Each component model is trained on data from a specific channel type (e.g., telephone, microphone).
  • Supervised Training: Crucially, the mixture is "supervised." This means the training data is segmented by channel type, allowing for the creation of dedicated PLDA models for each channel. This contrasts with unsupervised mixtures where the model learns the segmentation itself.
  • Scoring Fusion: The "mixture" aspect comes into play during the scoring phase. The likelihoods from each individual channel-specific PLDA model are combined (fused) to produce a final score. The paper uses a soft Bayesian fusion approach.
  • Channel-Independent PLDA (CI-PLDA): A specific component of the mixture is a PLDA trained on a combined dataset of both telephone and microphone data. This aims to create a model that is inherently less sensitive to channel differences.
  • i-vector Extractor: A universal i-vector extractor is employed, designed to handle both telephone and microphone speech effectively. This extractor uses separate total variability matrices for phone and mic data to address imbalances in dataset sizes.
  • Dimensionality Reduction: After i-vector extraction, Linear Discriminant Analysis (LDA) is used to reduce the dimensionality of the i-vectors, followed by normalization.

Datasets

The experiments were conducted using data from the NIST Speaker Recognition Evaluation (SRE) 2010 corpora. Specifically:

  • Training:
    • Telephone recordings from NIST SRE 2002/2003/2004/2005/2006/2008 (11256 recordings, 1250 male speakers).
  • Microphone-interview recordings from NIST SRE 2005/2006/2008 (4705 recordings, 203 male speakers).
  • Testing: The male part of the NIST SRE 2010 corpora was used for evaluation, specifically for:
  • Interview-interview (DET-2)
    • Phone-interview (DET-3)
    • Phone-phone (DET-5)

Key Insights and Ideas

  • Mixture Superiority: A mixture of channel-specific PLDA models significantly outperforms individual channel-specific models, especially in cross-channel scenarios. This highlights the benefit of leveraging diverse channel information.
  • Optimal Mixture Configuration: For cross-channel verification, mixing three channel-dependent PLDA systems (phone, mic, and a combined phone/mic) proved more effective than simpler mixtures.
  • Threshold Stability: The proposed mixture approach leads to a more stable decision threshold, which is crucial for practical applications and reliable cost assessment. This stability is a significant advantage over individual models.
  • Addressing Data Imbalance: The paper effectively addresses the common issue of imbalanced datasets between telephone and microphone recordings by using a universal i-vector extractor with separate total variability matrices and by carefully constructing the channel-independent component of the mixture.

Key Results

The experiments on the NIST SRE 2010 male dataset demonstrated the effectiveness of the supervised mixture of PLDA models:

  • Best Performance: The proposed "mix-PLDA" system achieved the best Equal Error Rate (EER) across most tests and the best normalized minimum detection cost function (min-DCF) in most cases.
  • Cross-Channel Improvement: In the phone-interview (DET-3) test, the mix-PLDA system showed a 3% EER reduction compared to the channel-independent CI-PLDA system.
  • Robustness: The mix-PLDA system exhibited good calibration and stable decision thresholds, even under unified conditions (common DET-2,3,5), outperforming individual channel-specific systems in this regard. For instance, the phone-PLDA system showed significant variability in decision thresholds under unified conditions.

Limitations and Future Work

  • Limited to Male Speakers: The experiments were conducted solely on the male portion of the NIST SRE 2010 dataset. The authors do not explicitly discuss the performance or applicability to female speakers, though the UBM used was gender-independent.
  • Future Work: The paper suggests that the observed stability of the decision threshold for the mixture model is significant for commercial applications, implying a direction for future research in deploying such systems. Further exploration of different mixture configurations and fusion strategies could also be a natural extension.

Summary updated

BibTeX
@inproceedings{simonchik2012supervized,
  author = {Konstantin Simonchik and Timur Pekhovsky and Andrey Shulipa and Anton Afanasyev},
  title = {{Supervized mixture of PLDA models for cross-channel speaker verification}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure